Better English for documenting when to use unimplemented!()
I don't think "plan of using" is correct here. I considered "plan on using" but eventually decided "plan to use" is better.
Bump bootstrap compiler to beta 1.53.0
This PR bumps the bootstrap compiler to version 1.53.0 beta, as part of our usual release process (this was supposed to be Wednesday's step, but creating the beta release took longer than expected).
The PR also includes the "Bootstrap: skip rustdoc fingerprint for building docs" commit, see the reasoning [on Zulip](https://zulip-archive.rust-lang.org/241545trelease/88450153betabootstrap.html).
r? `@Mark-Simulacrum`
Extend `rustc_on_implemented` to improve more `?` error messages
`_Self` could match the generic definition; this adds that functionality for matching the generic definition of type parameters too.
Your advice welcome on the wording of all these messages, and which things belong in the message/label/note.
r? `@estebank`
fix pad_integral example
pad_integral's parameter `is_nonnegative - whether the original integer was either positive or zero`, but in example it checked as `self.nb > 0`, so it previously printed `-0` for `format!("{}", Foo::new(0)`, what is wrong.
Extremely outdated; not only are traits implemented on arrays of arbitrary length, those implementations are documented on the primitive type, not in this module.
Rollup of 8 pull requests
Successful merges:
- #84717 (impl FromStr for proc_macro::Literal)
- #85169 (Add method-toggle to <details> for methods)
- #85287 (Expose `Concurrent` (private type in public i'face))
- #85315 (adding time complexity for partition_in_place iter method)
- #85439 (Add diagnostic item to `CStr`)
- #85464 (Fix UB in documented example for `ptr::swap`)
- #85470 (Fix invalid CSS rules for a:hover)
- #85472 (CTFE Machine: do not expose Allocation)
Failed merges:
r? `@ghost`
`@rustbot` modify labels: rollup
Avoid zero-length memcpy in formatting
This has two separate and somewhat orthogonal commits. The first change adjusts the ToString general impl for all types that implement Display; it no longer uses the full format machinery, rather directly falling onto a `std::fmt::Display::fmt` call. The second change directly adjusts the general core::fmt::write function which handles the production of format_args! to avoid zero-length push_str calls.
Both changes target the fact that push_str will still call memmove internally (or a similar function), as it doesn't know the length of the passed string. For zero-length strings in particular, this is quite expensive, and even for very short (several bytes long) strings, this is also expensive. Future work in this area may wish to have us fallback to write_char or similar, which may be cheaper on the (typically) short strings between the interpolated pieces in format_args!.
adding time complexity for partition_in_place iter method
I feel that one thing missing from rust docs compared to cpp references is existence of time complexity for all methods and functions. While it would be humongous task to include it for everything in single go, it is still doable if we as community keep on adding it in relevant places as and when we find them.
This PR adds the time complexity for partition_in_place method in iter.
remove InPlaceIterable marker from Peekable due to unsoundness
The unsoundness is not in Peekable per se, it rather is due to the
interaction between Peekable being able to hold an extra item
and vec::IntoIter's clone implementation shortening the allocation.
An alternative solution would be to change IntoIter's clone implementation
to keep enough spare capacity available.
fixes#85322
Override `clone_from` for some types
Override `clone_from` method of the `Clone` trait for:
- `cell::RefCell`
- `cmp::Reverse`
- `io::Cursor`
- `mem::ManuallyDrop`
This can bring performance improvements.
The unsoundness is not in Peekable per se, it rather is due to the
interaction between Peekable being able to hold an extra item
and vec::IntoIter's clone implementation shortening the allocation.
An alternative solution would be to change IntoIter's clone implementation
to keep enough spare capacity available.
Implement the new desugaring from `try_trait_v2`
~~Currently blocked on https://github.com/rust-lang/rust/issues/84782, which has a PR in https://github.com/rust-lang/rust/pull/84811~~ Rebased atop that fix.
`try_trait_v2` tracking issue: https://github.com/rust-lang/rust/issues/84277
Unfortunately this is already touching a ton of things, so if you have suggestions for good ways to split it up, I'd be happy to hear them. (The combination between the use in the library, the compiler changes, the corresponding diagnostic differences, even MIR tests mean that I don't really have a great plan for it other than trying to have decently-readable commits.
r? `@ghost`
~~(This probably shouldn't go in during the last week before the fork anyway.)~~ Fork happened.
Implement more Iterator methods on core::iter::Repeat
`core::iter::Repeat` always returns the same element, which means we can
do better than implementing most `Iterator` methods in terms of
`Iterator::next`.
Fixes#81292.
#81292 raises the question of whether these changes violate the contract of `core::iter::Repeat`, but as far as I can tell `core::iter::repeat` doesn't make any guarantees around how it calls `Clone::clone`.
# Stabilization report
## Summary
This stabilizes using macro expansion in key-value attributes, like so:
```rust
#[doc = include_str!("my_doc.md")]
struct S;
#[path = concat!(env!("OUT_DIR"), "/generated.rs")]
mod m;
```
See the changes to the reference for details on what macros are allowed;
see Petrochenkov's excellent blog post [on internals](https://internals.rust-lang.org/t/macro-expansion-points-in-attributes/11455)
for alternatives that were considered and rejected ("why accept no more
and no less?")
This has been available on nightly since 1.50 with no major issues.
## Notes
### Accepted syntax
The parser accepts arbitrary Rust expressions in this position, but any expression other than a macro invocation will ultimately lead to an error because it is not expected by the built-in expression forms (e.g., `#[doc]`). Note that decorators and the like may be able to observe other expression forms.
### Expansion ordering
Expansion of macro expressions in "inert" attributes occurs after decorators have executed, analogously to macro expressions appearing in the function body or other parts of decorator input.
There is currently no way for decorators to accept macros in key-value position if macro expansion must be performed before the decorator executes (if the macro can simply be copied into the output for later expansion, that can work).
## Test cases
- https://github.com/rust-lang/rust/blob/master/src/test/ui/attributes/key-value-expansion-on-mac.rs
- https://github.com/rust-lang/rust/blob/master/src/test/rustdoc/external-doc.rs
The feature has also been dogfooded extensively in the compiler and
standard library:
- https://github.com/rust-lang/rust/pull/83329
- https://github.com/rust-lang/rust/pull/83230
- https://github.com/rust-lang/rust/pull/82641
- https://github.com/rust-lang/rust/pull/80534
## Implementation history
- Initial proposal: https://github.com/rust-lang/rust/issues/55414#issuecomment-554005412
- Experiment to see how much code it would break: https://github.com/rust-lang/rust/pull/67121
- Preliminary work to restrict expansion that would conflict with this
feature: https://github.com/rust-lang/rust/pull/77271
- Initial implementation: https://github.com/rust-lang/rust/pull/78837
- Fix for an ICE: https://github.com/rust-lang/rust/pull/80563
## Unresolved Questions
~~https://github.com/rust-lang/rust/pull/83366#issuecomment-805180738 listed some concerns, but they have been resolved as of this final report.~~
## Additional Information
There are two workarounds that have a similar effect for `#[doc]`
attributes on nightly. One is to emulate this behavior by using a limited version of this feature that was stabilized for historical reasons:
```rust
macro_rules! forward_inner_docs {
($e:expr => $i:item) => {
#[doc = $e]
$i
};
}
forward_inner_docs!(include_str!("lib.rs") => struct S {});
```
This also works for other attributes (like `#[path = concat!(...)]`).
The other is to use `doc(include)`:
```rust
#![feature(external_doc)]
#[doc(include = "lib.rs")]
struct S {}
```
The first works, but is non-trivial for people to discover, and
difficult to read and maintain. The second is a strange special-case for
a particular use of the macro. This generalizes it to work for any use
case, not just including files.
I plan to remove `doc(include)` when this is stabilized. The
`forward_inner_docs` workaround will still compile without warnings, but
I expect it to be used less once it's no longer necessary.
This avoids a zero-length write_str call, which boils down to a zero-length
memmove and ultimately costs quite a few instructions on some workloads.
This is approximately a 0.33% instruction count win on diesel-check.
`core::iter::Repeat` always returns the same element, which means we can
do better than implementing most `Iterator` methods in terms of
`Iterator::next`.
Fixes#81292.
add BITS associated constant to core::num::Wrapping
This keeps `Wrapping` synchronized with the primitives it wraps as for the #32463 `wrapping_int_impl` feature.
#[inline(always)] on basic pointer methods
Retryng #85201 with only inlining pointer methods. The goal is to make pointers behave just like pointers in O0, mainly to reduce overhead in debug builds.
cc `@scottmcm`
Add auto traits and clone trait migrations for RFC2229
This PR
- renames the existent RFC2229 migration `disjoint_capture_drop_reorder` to `disjoint_capture_migration`
- add additional migrations for auto traits and clone trait
Closesrust-lang/project-rfc-2229#29Closesrust-lang/project-rfc-2229#28
r? `@nikomatsakis`
Make unchecked_{add,sub,mul} inherent methods unstably const
The intrinsics are marked as being stably const (even though they're not stable by nature of being intrinsics), but the currently-unstable inherent versions are not marked as const. This fixes this inconsistency. Split out of #85017,
r? `@oli-obk`
Disallows `#![feature(no_coverage)]` on stable and beta (using standard crate-level gating)
Fixes: #84836
Removes the function-level feature gating solution originally implemented, and solves the same problem using `allow_internal_unstable`, so normal crate-level feature gating mechanism can still be used (which disallows the feature on stable and beta).
I tested this, building the compiler with and without `CFG_DISABLE_UNSTABLE_FEATURES=1`
With unstable features disabled, I get the expected result as shown here:
```shell
$ ./build/x86_64-unknown-linux-gnu/stage1/bin/rustc src/test/run-make-fulldeps/coverage/no_cov_crate.rs
error[E0554]: `#![feature]` may not be used on the dev release channel
--> src/test/run-make-fulldeps/coverage/no_cov_crate.rs:2:1
|
2 | #![feature(no_coverage)]
| ^^^^^^^^^^^^^^^^^^^^^^^^
error: aborting due to previous error
For more information about this error, try `rustc --explain E0554`.
```
r? ````@Mark-Simulacrum````
cc: ````@tmandry```` ````@wesleywiser````
Allow using `core::` in intra-doc links within core itself
I came up with this idea ages ago, but rustdoc used to ICE on it. Now it doesn't.
Helps with https://github.com/rust-lang/rust/issues/73445. Doesn't fix it completely since `extern crate self as std;` in std still gives strange errors.
Clarify documentation for `[T]::contains`
Change the documentation to correctly characterize when the suggested alternative to `contains` applies, and correctly explain why it works.
Fixes#84877
Update `ptr` docs with regards to `ptr::addr_of!`
This updates the documentation since `ptr::addr_of!` and `ptr::addr_of_mut!` are now stable. One might remove the distinction between the sections `# On packed structs` and `# Examples`, as the old section on packed structs was primarily to prevent users of doing undefined behavior, which is not necessary anymore.
Technically there is now wrong/outdated documentation on stable, but I don't think this is worth a point release 😉Fixes#83509.
``````````@rustbot`````````` modify labels: T-doc
using allow_internal_unstable (as recommended)
Fixes: #84836
```shell
$ ./build/x86_64-unknown-linux-gnu/stage1/bin/rustc src/test/run-make-fulldeps/coverage/no_cov_crate.rs
error[E0554]: `#![feature]` may not be used on the dev release channel
--> src/test/run-make-fulldeps/coverage/no_cov_crate.rs:2:1
|
2 | #![feature(no_coverage)]
| ^^^^^^^^^^^^^^^^^^^^^^^^
error: aborting due to previous error
For more information about this error, try `rustc --explain E0554`.
```
This updates the documentation since `ptr::addr_of!` and
`ptr::addr_of_mut!` are now stable. One might remove the distinction
between the sections `# On packed structs` and `# Examples`, as the old
section on packed structs was primarily to prevent users of doing unde-
fined behavior, which is not necessary anymore.
There is also a new section in "how to obtain a pointer", which referen-
ces the `ptr::addr_of!` macros.
This commit contains squashed commits from code review.
Co-authored-by: Joshua Nelson <joshua@yottadb.com>
Co-authored-by: Mara Bos <m-ou.se@m-ou.se>
Co-authored-by: Soveu <marx.tomasz@gmail.com>
Co-authored-by: Ralf Jung <post@ralfj.de>
[Arm64] use isb instruction instead of yield in spin loops
On arm64 we have seen on several databases that ISB (instruction synchronization
barrier) is better to use than yield in a spin loop. The yield instruction is a
nop. The isb instruction puts the processor to sleep for some short time. isb
is a good equivalent to the pause instruction on x86.
Below is an experiment that shows the effects of yield and isb on Arm64 and the
time of a pause instruction on x86 Intel processors. The micro-benchmarks use
https://github.com/google/benchmark.git
```
$ cat a.cc
static void BM_scalar_increment(benchmark::State& state) {
int i = 0;
for (auto _ : state)
benchmark::DoNotOptimize(i++);
}
BENCHMARK(BM_scalar_increment);
static void BM_yield(benchmark::State& state) {
for (auto _ : state)
asm volatile("yield"::);
}
BENCHMARK(BM_yield);
static void BM_isb(benchmark::State& state) {
for (auto _ : state)
asm volatile("isb"::);
}
BENCHMARK(BM_isb);
BENCHMARK_MAIN();
$ g++ -o run a.cc -O2 -lbenchmark -lpthread
$ ./run
--------------------------------------------------------------
Benchmark Time CPU Iterations
--------------------------------------------------------------
AWS Graviton2 (Neoverse-N1) processor:
BM_scalar_increment 0.485 ns 0.485 ns 1000000000
BM_yield 0.400 ns 0.400 ns 1000000000
BM_isb 13.2 ns 13.2 ns 52993304
AWS Graviton (A-72) processor:
BM_scalar_increment 0.897 ns 0.874 ns 801558633
BM_yield 0.877 ns 0.875 ns 800002377
BM_isb 13.0 ns 12.7 ns 55169412
Apple Arm64 M1 processor:
BM_scalar_increment 0.315 ns 0.315 ns 1000000000
BM_yield 0.313 ns 0.313 ns 1000000000
BM_isb 9.06 ns 9.06 ns 77259282
```
```
static void BM_pause(benchmark::State& state) {
for (auto _ : state)
asm volatile("pause"::);
}
BENCHMARK(BM_pause);
Intel Skylake processor:
BM_scalar_increment 0.295 ns 0.295 ns 1000000000
BM_pause 41.7 ns 41.7 ns 16780553
```
Tested on Graviton2 aarch64-linux with `./x.py test`.
On arm64 we have seen on several databases that ISB (instruction synchronization
barrier) is better to use than yield in a spin loop. The yield instruction is a
nop. The isb instruction puts the processor to sleep for some short time. isb
is a good equivalent to the pause instruction on x86.
Below is an experiment that shows the effects of yield and isb on Arm64 and the
time of a pause instruction on x86 Intel processors. The micro-benchmarks use
https://github.com/google/benchmark.git
$ cat a.cc
static void BM_scalar_increment(benchmark::State& state) {
int i = 0;
for (auto _ : state)
benchmark::DoNotOptimize(i++);
}
BENCHMARK(BM_scalar_increment);
static void BM_yield(benchmark::State& state) {
for (auto _ : state)
asm volatile("yield"::);
}
BENCHMARK(BM_yield);
static void BM_isb(benchmark::State& state) {
for (auto _ : state)
asm volatile("isb"::);
}
BENCHMARK(BM_isb);
BENCHMARK_MAIN();
$ g++ -o run a.cc -O2 -lbenchmark -lpthread
$ ./run
--------------------------------------------------------------
Benchmark Time CPU Iterations
--------------------------------------------------------------
AWS Graviton2 (Neoverse-N1) processor:
BM_scalar_increment 0.485 ns 0.485 ns 1000000000
BM_yield 0.400 ns 0.400 ns 1000000000
BM_isb 13.2 ns 13.2 ns 52993304
AWS Graviton (A-72) processor:
BM_scalar_increment 0.897 ns 0.874 ns 801558633
BM_yield 0.877 ns 0.875 ns 800002377
BM_isb 13.0 ns 12.7 ns 55169412
Apple Arm64 M1 processor:
BM_scalar_increment 0.315 ns 0.315 ns 1000000000
BM_yield 0.313 ns 0.313 ns 1000000000
BM_isb 9.06 ns 9.06 ns 77259282
static void BM_pause(benchmark::State& state) {
for (auto _ : state)
asm volatile("pause"::);
}
BENCHMARK(BM_pause);
Intel Skylake processor:
BM_scalar_increment 0.295 ns 0.295 ns 1000000000
BM_pause 41.7 ns 41.7 ns 16780553
Tested on Graviton2 aarch64-linux with `./x.py test`.
Searching for "reduce" currently puts the `reduce` alias for `fold`
above the actual `reduce` function. The `reduce` function already has a
cross-reference for `fold`, and vice versa.
The Eq trait has a special hidden function. MIR `InstrumentCoverage`
would add this function to the coverage map, but it is never called, so
the `Eq` trait would always appear uncovered.
Fixes: #83601
The fix required creating a new function attribute `no_coverage` to mark
functions that should be ignored by `InstrumentCoverage` and the
coverage `mapgen` (during codegen).
While testing, I also noticed two other issues:
* spanview debug file output ICEd on a function with no body. The
workaround for this is included in this PR.
* `assert_*!()` macro coverage can appear covered if followed by another
`assert_*!()` macro. Normally they appear uncovered. I submitted a new
Issue #84561, and added a coverage test to demonstrate this issue.
Add the `try_trait_v2` library basics
No compiler changes as part of this -- just new unstable traits and impls thereof.
The goal here is to add the things that aren't going to break anything, to keep the feature implementation simpler in the next PR.
(Draft since the FCP won't end until Saturday, but I was feeling optimistic today -- and had forgotten that FCP was 10 days, not 7 days.)
Stabilize Duration::MAX
Following the suggested direction from https://github.com/rust-lang/rust/issues/76416#issuecomment-817278338, this PR proposes that `Duration::MAX` should have been part of the `duration_saturating_ops` feature flag all along, having been
0. heavily referenced by that feature flag
1. an odd duck next to most of `duration_constants`, as I expressed in https://github.com/rust-lang/rust/issues/57391#issuecomment-717681193
2. introduced in #76114 which added `duration_saturating_ops`
and accordingly should be folded into `duration_saturating_ops` and therefore stabilized.
r? `@m-ou-se`
move core::hint::black_box under its own feature gate
The `black_box` function had its own RFC and is tracked separately from the `test` feature at https://github.com/rust-lang/rust/issues/64102. Let's reflect this in the feature gate.
To avoid breaking all the benchmarks, libtest's `test::black_box` is a wrapping definition, not a reexport -- this means it is still under the `test` feature gate.
Cautiously add IntoIterator for arrays by value
Add the attribute described in #84133, `#[rustc_skip_array_during_method_dispatch]`, which effectively hides a trait from method dispatch when the receiver type is an array.
Then cherry-pick `IntoIterator for [T; N]` from #65819 and gate it with that attribute. Arrays can now be used as `IntoIterator` normally, but `array.into_iter()` has edition-dependent behavior, returning `slice::Iter` for 2015 and 2018 editions, or `array::IntoIter` for 2021 and later.
r? `@nikomatsakis`
cc `@LukasKalbertodt` `@rust-lang/libs`
Duration is used in std to represent a difference between two Instants.
As such, it has to at least contain that span of time in it. However,
Instant can vary by platform. Thus, we should explain the impl of
Duration::MAX is sensitive to these vagaries of the platform.
further split up const_fn feature flag
This continues the work on splitting up `const_fn` into separate feature flags:
* `const_fn_trait_bound` for `const fn` with trait bounds
* `const_fn_unsize` for unsizing coercions in `const fn` (looks like only `dyn` unsizing is still guarded here)
I don't know if there are even any things left that `const_fn` guards... at least libcore and liballoc do not need it any more.
`@oli-obk` are you currently able to do reviews?
Mention FusedIterator case in Iterator::fuse doc
Using `fuse` on an iterator that incorrectly implements
`FusedIterator` does not fuse the iterator. This commit adds a
note about this in the documentation of this method to increase
awareness about this potential issue (esp. when relying on fuse
in unsafe code).
Closes#83969
implement `TrustedRandomAccess` for `Take` iterator adapter
`TrustedRandomAccess` requires the iterator length to fit within `usize`. `take(n)` only constrains the upper bound of an iterator. So if the inner is `TrustedRandomAccess` (which already implies a finite length) then so can be `Take`.
```````@rustbot``````` label T-libs-impl
Improve `Iterator::by_ref` example
I split the example into two: one that fails to compile, and one that
works. I also made them identical except for the addition of `by_ref`
so we don't confuse readers with random differences.
cc `@steveklabnik,` who is the one that added the previous version of this example
Using `fuse` on an iterator that incorrectly implements
`FusedIterator` does not fuse the iterator. This commit adds a
note about this in the documentation of this method to increase
awareness about this potential issue (esp. when relying on fuse
in unsafe code).
Added CharIndices::offset function
The CharIndices iterator has a field internally called front_offset, that I think would be very useful to have access to.
You can already do something like ``char_indices.next().map(|(offset, _)| offset)``, but that is wordy, in addition to not handling the case where the iterator has ended, where you'd want the offset to be equal to the length.
I'm very new to the open source world and the rust repository, so I'm sorry if I missed a step or did something weird.
Implement indexing slices with pairs of core::ops::Bound<usize>
Closes#49976.
I am not sure about code duplication between `check_range` and `into_maybe_range`. Should be former implemented in terms of the latter? Also this PR doesn't address code duplication between `impl SliceIndex for Range*`.
Format `Struct { .. }` on one line even with `{:#?}`.
The result of `debug_struct("A").finish_non_exhaustive()` before this change:
```
A {
..
}
```
And after this change:
```
A { .. }
```
If there's any fields, the result stays unchanged:
```
A {
field: value,
..
}
fix 'const-stable since' for NonZeroU*::new_unchecked
For the unsigned `NonZero` types, `new_unchecked` was const-stable from the start with https://github.com/rust-lang/rust/pull/50808. Fix the docs to accurately reflect that.
I think this `since` is also incorrect:
```rust
#[stable(feature = "from_nonzero", since = "1.31.0")]
impl From<$Ty> for $Int {
```
The signed nonzero types were only stabilized in 1.34, so that `From` impl certainly didn't exist before. But I had enough of digging through git histories after I figured out when `new_unchecked` became const-stable...^^
Replace all `fmt.pad` with `debug_struct`
This replaces any occurrence of:
- `f.pad("X")` with `f.debug_struct("X").finish()`
- `f.pad("X { .. }")` with `f.debug_struct("X").finish_non_exhaustive()`
This is in line with existing formatting code such as
1255053067/library/std/src/sync/mpsc/mod.rs (L1470-L1475)
Deprecate the core::raw / std::raw module
It only contains the `TraitObject` struct which exposes components of wide pointer. Pointer metadata APIs are designed to replace this: https://github.com/rust-lang/rust/issues/81513
Add some #[inline(always)] to arithmetic methods of integers
I tried to add it only to methods which return results of intrinsics and don't have any branching.
Branching could made performance of debug builds (`-Copt-level=0`) worse.
Main goal of changes is allowing wider optimizations in `-Copt-level=1`.
Closes: https://github.com/rust-lang/rust/issues/75598
r? `@nagisa`
No compiler changes as part of this -- just new unstable traits and impls thereof.
The goal here is to add the things that aren't going to break anything, to keep the feature implementation simpler in the next PR.
add lint deref_nullptr detecting when a null ptr is dereferenced
fixes#83856
changelog: add lint that detect code like
```rust
unsafe {
&*core::ptr::null::<i32>()
};
unsafe {
addr_of!(std::ptr::null::<i32>())
};
let x: i32 = unsafe {*core::ptr::null()};
let x: i32 = unsafe {*core::ptr::null_mut()};
unsafe {*(0 as *const i32)};
unsafe {*(core::ptr::null() as *const i32)};
```
```
warning: Dereferencing a null pointer causes undefined behavior
--> src\main.rs:5:26
|
5 | let x: i32 = unsafe {*core::ptr::null()};
| ^^^^^^^^^^^^^^^^^^
| |
| a null pointer is dereferenced
| this code causes undefined behavior when executed
|
= note: `#[warn(deref_nullptr)]` on by default
```
Limitation:
It does not detect code like
```rust
const ZERO: usize = 0;
unsafe {*(ZERO as *const i32)};
```
or code where `0` is not directly a literal
I tried to add it only to methods which return results of intrinsics and don't have any branching.
Branching could made performance of debug builds (`-Copt-level=0`) worse.
Main goal of changes is allowing wider optimizations in `-Copt-level=1`.
Closes: https://github.com/rust-lang/rust/issues/75598
Add note about reverting a workaround in the future
The root cause was fixed upstream in LLVM main. This adds a reminder to revert the workaround once the LLVM rustc depends on is new enough. Since I'm not sure how such optimizations get routed to LLVM releases, I used the conservative assumption that it will only show up with LLVM 13.
